Skip to content
Hot storiesContinuously updated

研究评估4位微缩放格式在可穿戴模型上的表现

1 reports1 reporting sourcesUpdated 9 hours ago

Start here

AI overview

2026年11月4日,arXiv发布针对可穿戴动作识别模型的研究。研究人员在包含159,492个参数的可穿戴1D CNN模型上评估了4-bit微缩放格式,当权重与激活值均量化为4-bit时,NVFP4表现显著优于MXFP4。在全精度macro-F1基准为69.26%的测试中,NVFP4仅损失0.75个百分点,而MXFP4损失达5.14个百分点;在5折验证中,NVFP4持续领先MXFP4达3.49至4.40个百分点。消融实验进一步证实,缩放因子下取整是导致MXFP4精度损失的主要原因,改用上取整可挽回超过3个百分点的精度损失。

AI-generated from reports · Updated 9 hours ago

Report timeline

Follow the reports to see different sides of the story.

Nov 4
  1. arXiv 机器学习
    4-Bit Microscaling 会给可穿戴动作识别模型带来多大精度损失?

    针对 159,492 参数可穿戴 1D CNN 模型,权重与激活值均量化为 4-bit 时,NVFP4 仅损失 0.75 个百分点 macro-F1,远优于损失 5.14 点的 MXFP4。在 69.26% 全精度基准下,NVFP4 在 5 折验证中持续领先 MXFP4 达 3.49 至 4.40 点;消融实验证实缩放因子下取整是 MXFP4 损失主因,改用上取整可挽回超 3 个百分点。

Story heat trend

Current heat 8·Comparable peak 10(Nov 4 14:00)·Comparable change, last 24 hours–

02.557.510Nov4Nov4Nov4Nov4

Trends compare only the same parties observed continuously, so the scope can be narrower than current heat counts. Move the pointer or tap the chart to see hourly heat; use the left and right arrow keys on a keyboard.